Brief Review — Automatic liver tumor segmentation in CT with fully convolutional neural networks and object-based postprocessing

U-Net as Segmentation Network + RF as Post Processing

Sik-Ho Tsang
3 min readJan 24, 2023

Automatic liver tumor segmentation in CT with fully convolutional neural networks and object-based postprocessing,
FCN+RF, by Fraunhofer Institute for Medical Image Computing MEVIS, Radboud University Medical Center, Jacobs University, and University of Bremen,
2018 Nature Sci. Rep., Over 170 Citations (

@ Medium)
Medical Image Analysis, Medical Imaging, Image Segmentation, U-Net

  • A fully automatic method for liver tumor segmentation in CT images is developed based on a 2D fully convolutional neural network with an object-based postprocessing step.

Outline

  1. U-Net+RF
  2. Results

1. U-Net+RF

1.1. Segmentation Model Using U-Net

Overview of the neural network architecture. The numbers denote the feature map count.
  • U-Net like model is used, which works on four resolution levels allowing for learning of local and global features.
  • The network contains long skip connections, and also short skip connections.
  • Each convolutional layer uses 3×3 filter size and is followed by a batch normalization and a ReLU activation function.
  • Dropout (p = 0.5) is used before each convolution in the upscaling path to prevent the network from overfitting.

1.2. Object-based Postprocessing Using RF

  • Based on the training data we observed that some neural network outputs corresponded to false positives, which could easily be identified by their shape and location.
  • A post-processing step is added, which employs a model classifying tumor objects (computed as 3D connected components of the FCN output) into true (TP) and false positives (FP).
  • A conventional random forest classifier (RF) is trained, with 256 trees using 36 hand-crafted features carrying information about underlying image statistics, tumor shape and its distance to the liver boundary.
  • Because they work well with moderate numbers of training samples and varying feature value distributions.

2. Results

Mean metric values for human vs. human and computer vs. human comparisons.
Box plots showing dice per case (a) an dice per correspondence (b) computed for expert and automatically generated segmentations on 30 test cases.
  • LiTS: Ground-truth labels provided by dataset.
  • MTRA: An expert asked by authors to annotate again.
  • The neural network was able to detect 47% and 72% of all tumors present in the MTRA and LiTS annotations, respectively.
  • The dice/case and dice/correspondence was 0.53 and 0.72 for the MTRA reference and 0.51 and 0.65 for LiTS reference.
  • The RF classifier allowed for a 85% reduction of false positives and had 87% accuracy on test cases. The improvement for Dice per correspondence was significant.
  • At that moment, it is ranking third at the MICCAI 2017 LiTS round (leaderboard user name hans.meine). The submission scored 0.68 and 0.96 dice/case for tumor and liver segmentation, respectively.
  • The proposed method needs on average 67s for one case: 43, 16 and 8s for liver segmentation, tumor segmentation and FP filtering, respectively.
MTRA (dashed) vs. LiTS (solid) annotations.
Neural network (black) compared with the LiTS (white) annotations.
  • Some visualizations are shown above.

--

--

Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: https://linktr.ee/shtsang for Twitter, LinkedIn, etc.